Convolutional Neural Networks¶
Project: Write an Algorithm for a Dog Identification App¶
In this notebook, some template code has already been provided for you, and you will need to implement additional functionality to successfully complete this project. You will not need to modify the included code beyond what is requested. Sections that begin with '(IMPLEMENTATION)' in the header indicate that the following block of code will require additional functionality which you must provide. Instructions will be provided for each section, and the specifics of the implementation are marked in the code block with a 'TODO' statement. Please be sure to read the instructions carefully!
Note: Once you have completed all of the code implementations, you need to finalize your work by exporting the Jupyter Notebook as an HTML document. Before exporting the notebook to html, all of the code cells need to have been run so that reviewers can see the final implementation and output. You can then export the notebook by using the menu above and navigating to File -> Download as -> HTML (.html). Include the finished document along with this notebook as your submission.
In addition to implementing code, there will be questions that you must answer which relate to the project and your implementation. Each section where you will answer a question is preceded by a 'Question X' header. Carefully read each question and provide thorough answers in the following text boxes that begin with 'Answer:'. Your project submission will be evaluated based on your answers to each of the questions and the implementation you provide.
Note: Code and Markdown cells can be executed using the Shift + Enter keyboard shortcut. Markdown cells can be edited by double-clicking the cell to enter edit mode.
The rubric contains optional "Stand Out Suggestions" for enhancing the project beyond the minimum requirements. If you decide to pursue the "Stand Out Suggestions", you should include the code in this Jupyter notebook.
Why We're Here¶
In this notebook, you will make the first steps towards developing an algorithm that could be used as part of a mobile or web app. At the end of this project, your code will accept any user-supplied image as input. If a dog is detected in the image, it will provide an estimate of the dog's breed. If a human is detected, it will provide an estimate of the dog breed that is most resembling. The image below displays potential sample output of your finished project (... but we expect that each student's algorithm will behave differently!).
In this real-world setting, you will need to piece together a series of models to perform different tasks; for instance, the algorithm that detects humans in an image will be different from the CNN that infers dog breed. There are many points of possible failure, and no perfect algorithm exists. Your imperfect solution will nonetheless create a fun user experience!
The Road Ahead¶
We break the notebook into separate steps. Feel free to use the links below to navigate the notebook.
- Step 0: Import Datasets
- Step 1: Detect Humans
- Step 2: Detect Dogs
- Step 3: Create a CNN to Classify Dog Breeds (from Scratch)
- Step 4: Create a CNN to Classify Dog Breeds (using Transfer Learning)
- Step 5: Write your Algorithm
- Step 6: Test Your Algorithm
Step 0: Import Datasets¶
Make sure that you've downloaded the required human and dog datasets:
Download the dog dataset. Unzip the folder and place it in this project's home directory, at the location
/dog_images.Download the human dataset. Unzip the folder and place it in the home directory, at location
/lfw.
Note: If you are using a Windows machine, you are encouraged to use 7zip to extract the folder.
In the code cell below, we save the file paths for both the human (LFW) dataset and dog dataset in the numpy arrays human_files and dog_files.
from ipywidgets import *
from IPython.display import display
from widget_helper import *
import pandas as pd
from IPython.display import display, HTML
from tqdm import tqdm
import numpy as np
from glob import glob
from enum import Enum
import torch.nn as nn
import torch.nn.functional as F
import os
import torch
from torchvision import datasets,transforms
from PIL import ImageFile
import math
import random
import helpers as hlp
import models as mds
import time
import copy
# load filenames for human and dog images
human_files = np.array(glob("data/lfw/*/*"))
dog_files = np.array(glob("data/dogImages_crop_350/*/*/*"))
# print number of images in each dataset
print('There are %d total human images.' % len(human_files))
print('There are %d total dog images.' % len(dog_files))
Step 1: Detect Humans¶
In this section, we use OpenCV's implementation of Haar feature-based cascade classifiers to detect human faces in images.
OpenCV provides many pre-trained face detectors, stored as XML files on github. We have downloaded one of these detectors and stored it in the haarcascades directory. In the next code cell, we demonstrate how to use this detector to find human faces in a sample image.
import cv2
import matplotlib.pyplot as plt
%matplotlib inline
# extract pre-trained face detector
face_cascade = cv2.CascadeClassifier('haarcascades/haarcascade_frontalface_alt.xml')
# load color (BGR) image
filename = human_files[random.randint(0,len(human_files)-1)]
img = cv2.imread(filename)
# convert BGR image to grayscale
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# find faces in image
faces = face_cascade.detectMultiScale(gray)
# print number of faces detected in the image
print('Number of faces detected:', len(faces))
# get bounding box for each detected face
for (x,y,w,h) in faces:
# add bounding box to color image
cv2.rectangle(img,(x,y),(x+w,y+h),(255,0,0),2)
# convert BGR image to RGB for plotting
cv_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
# display the image, along with bounding box
plt.imshow(cv_rgb)
name = hlp.get_name_from_filename(filename)
plt.title(name)
plt.show()
samples = random.sample(range(0, len(human_files)-1), 5)
sample_files = human_files[samples]
hlp.show_images(image_files = sample_files, n_col = 5, image_size = 200, h = 20, w = 20)
samples = random.sample(range(0, len(dog_files)-1), 5)
sample_files = dog_files[samples]
hlp.show_images(image_files = sample_files, n_col = 5, image_size = 200, h = 20, w = 20)
Before using any of the face detectors, it is standard procedure to convert the images to grayscale. The detectMultiScale function executes the classifier stored in face_cascade and takes the grayscale image as a parameter.
In the above code, faces is a numpy array of detected faces, where each row corresponds to a detected face. Each detected face is a 1D array with four entries that specifies the bounding box of the detected face. The first two entries in the array (extracted in the above code as x and y) specify the horizontal and vertical positions of the top left corner of the bounding box. The last two entries in the array (extracted here as w and h) specify the width and height of the box.
Write a Human Face Detector¶
We can use this procedure to write a function that returns True if a human face is detected in an image and False otherwise. This function, aptly named face_detector, takes a string-valued file path to an image as input and appears in the code block below.
# returns "True" if face is detected in image stored at img_path
def face_detector(img_path):
img = cv2.imread(img_path)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
faces = face_cascade.detectMultiScale(gray)
return len(faces) > 0
(IMPLEMENTATION) Assess the Human Face Detector¶
Question 1: Use the code cell below to test the performance of the face_detector function.
- What percentage of the first 100 images in
human_fileshave a detected human face? - What percentage of the first 100 images in
dog_fileshave a detected human face?
Ideally, we would like 100% of human images with a detected face and 0% of dog images with a detected face. You will see that our algorithm falls short of this goal, but still gives acceptable performance. We extract the file paths for the first 100 images from each of the datasets and store them in the numpy arrays human_files_short and dog_files_short.
Answer:
99.0%human faces detected using face-detector with 1 exception8.0%dogs detected with human faces using face-detector
| Accuracy | F1-score |
|---|---|
| 0.925 | 0.93 |
human_files_short = human_files[:100]
dog_files_short = dog_files[:100]
#-#-# Do NOT modify the code above this line. #-#-#
## on the images in human_files_short and dog_files_short.
def display_performance(TP, FN, TN, FP):
# computing acccuracy and f1-score
precision=TP/(FP+TP)
recall=TP/(FN+TP)
accuracy=(TP+TN)/(TP+TN+FP+FN)
f1score=2*((precision*recall)/(precision+recall))
print('---------------------')
print('| Accuracy | {}|'.format(str(round(accuracy,3 ) ).rjust(7) ))
print('---------------------')
print('| F1-score | {}|'.format(str(round(f1score,3 ) ).rjust(7) ))
print('---------------------')
def eval_detector(sample_files, detector, positive):
"""
parameters:
sample_files: list of image file names to be detected
detector: Is a detector function i.e can be face detector or dog detector
"""
positive_files = []
negative_files = []
for file_path in tqdm(sample_files):
detected = detector(file_path)
if detected:
positive_files.append(file_path)
else:
negative_files.append(file_path)
if positive:
TP = len(positive_files)
FN = len(negative_files)
TN = None
FP = None
else:
TP = None
FN = None
TN = len(negative_files)
FP = len(positive_files)
return positive_files, negative_files, TP, FN, TN, FP
Evaluate human face detector¶
# detect human face in human_files_short images
human_face, no_human_face, TP, FN,_,_ = eval_detector(human_files_short, face_detector, positive=True)
print('-'*50)
print('{:.3f}% human faces detected with {} exception'.format((TP)/(FN+TP)*100, FN))
print('-'*50)
hlp.show_images(image_files = no_human_face, n_col = 4, image_size = 50, h = 20, w = 20)
plt.show();
# detect human face in dog_files_short images
human_face_in_dog, no_human_face_in_dog, _, _, TN, FP = eval_detector(dog_files_short, face_detector, positive=False)
print('-'*50)
print('{:.3f}% dogs detected with human faces'.format(((FP)/(FP+TN)*100)))
print('-'*50)
hlp.show_images(image_files = human_face_in_dog, n_col = 4, image_size = 50, h = 20, w = 20)
plt.show();
display_performance(TP, FN, TN, FP)
We suggest the face detector from OpenCV as a potential way to detect human images in your algorithm, but you are free to explore other approaches, especially approaches that make use of deep learning :). Please use the code cell below to design and test your own face detection algorithm. If you decide to pursue this optional task, report performance on human_files_short and dog_files_short.
### (Optional)
### TODO: Test performance of anotherface detection algorithm.
### Feel free to use as many code cells as needed.
Step 2: Detect Dogs¶
In this section, we use a pre-trained model to detect dogs in images.
Obtain Pre-trained VGG-16 Model¶
The code cell below downloads the VGG-16 model, along with weights that have been trained on ImageNet, a very large, very popular dataset used for image classification and other vision tasks. ImageNet contains over 10 million URLs, each linking to an image containing an object from one of 1000 categories.
import torch
import torchvision.models as models
# define VGG16 model
VGG16 = models.vgg16(pretrained=True)
# check if CUDA is available
use_cuda = torch.cuda.is_available()
# move model to GPU if CUDA is available
if use_cuda:
VGG16 = VGG16.cuda()
Given an image, this pre-trained VGG-16 model returns a prediction (derived from the 1000 possible categories in ImageNet) for the object that is contained in the image.
(IMPLEMENTATION) Making Predictions with a Pre-trained Model¶
In the next code cell, you will write a function that accepts a path to an image (such as 'dogImages/train/001.Affenpinscher/Affenpinscher_00001.jpg') as input and returns the index corresponding to the ImageNet class that is predicted by the pre-trained VGG-16 model. The output should always be an integer between 0 and 999, inclusive.
Before writing the function, make sure that you take the time to learn how to appropriately pre-process tensors for pre-trained models in the PyTorch documentation.
from PIL import Image
import torchvision.transforms as transforms
import json
def load_image(img_path, size=[224, 224]):
image = Image.open(img_path).convert('RGB')
in_transform = transforms.Compose([
transforms.Resize(size),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406),
(0.229, 0.224, 0.225))])
# discard the transparent, alpha channel (that's the :3) and add the batch dimension
image = in_transform(image)[:3,:,:].unsqueeze(0)
return image
def VGG16_predict(img_path):
'''
Use pre-trained VGG-16 model to obtain index corresponding to
predicted ImageNet class for image at specified path
Args:
img_path: path to an image
Returns:
Index corresponding to VGG-16 model's prediction
'''
## TODO: Complete the function.
## Load and pre-process an image from the given img_path
## Return the *index* of the predicted class for that image
#load and transform image
Image = load_image(img_path)
#predict using pre-trained VGG16 model
prediction = VGG16(Image)
#Get index
index = prediction.data.numpy().argmax()
return index # predicted class index
def get_imagenet_class_label(index):
class_idx = json.load(open("imagenet_class_index.json"))
idx2label = [class_idx[str(k)][1] for k in range(len(class_idx))]
return idx2label[index]
human_idx = random.randint(0, len(human_files)-1)
index_human = VGG16_predict(human_files[human_idx])
dog_idx = random.randint(0, len(dog_files)-1)
index_dog = VGG16_predict(dog_files[dog_idx])
sample_files = [human_files[human_idx]]
titles=['Imagenet Label: {} Index: {}'.format(get_imagenet_class_label(index_human), index_human)]
hlp.show_images(image_files = sample_files, n_col = 5, image_size = 200, h = 20, w = 20, titles=titles)
plt.show()
sample_files = [dog_files[dog_idx]]
titles=['Imagenet Label: {} Index: {}'.format(get_imagenet_class_label(index_dog), index_dog)]
hlp.show_images(image_files = sample_files, n_col = 5, image_size = 200, h = 20, w = 20, titles=titles)
plt.show()
def predict_imagenet_class_label(img_path):
imagenet_class_index = VGG16_predict(img_path)
return get_imagenet_class_label(imagenet_class_index)
(IMPLEMENTATION) Write a Dog Detector¶
While looking at the dictionary, you will notice that the categories corresponding to dogs appear in an uninterrupted sequence and correspond to dictionary keys 151-268, inclusive, to include all categories from 'Chihuahua' to 'Mexican hairless'. Thus, in order to check to see if an image is predicted to contain a dog by the pre-trained VGG-16 model, we need only check if the pre-trained model predicts an index between 151 and 268 (inclusive).
Use these ideas to complete the dog_detector function below, which returns True if a dog is detected in an image (and False if not).
### returns "True" if a dog is detected in the image stored at img_path
def dog_detector(img_path):
## TODO: Complete the function.
prediction = VGG16_predict(img_path)
return prediction>=151 and prediction<=268
print('Test- human_files[0] prediction: ',dog_detector(human_files[0]))
print('Test- dog_files[0] prediction: ',dog_detector(dog_files[0]))
(IMPLEMENTATION) Assess the Dog Detector¶
Question 2: Use the code cell below to test the performance of your dog_detector function.
- What percentage of the images in
human_files_shorthave a detected dog? - What percentage of the images in
dog_files_shorthave a detected dog?
Answer:
2.0%dogs detected in human faces using VGG-16 dog-detector100.0%dogs correctly detected with 0 exception using VGG-16 dog-detector
| Accuracy | F1-score |
|---|---|
| 0.995 | 0.995 |
### on the images in human_files_short and dog_files_short.
# Detect dogs in dog_files_short
dogs_detected, dogs_not_detected, TP, FN,_,_ = eval_detector(dog_files_short, dog_detector, positive=True)
print('-'*50)
print('{}% dogs correctly detected with {} exception'.format((TP)/(FN+TP)*100, FN))
print('-'*50)
hlp.show_images(image_files = dogs_not_detected, n_col = 4, image_size = 50, h = 10, w = 10)
plt.show();
# Detect dogs in human_files_short
dogs_detected_in_faces, no_dogs_detected_in_faces, _, _, TN, FP = eval_detector(human_files_short, dog_detector, positive=False)
print('-'*50)
print('{}% dogs detected in human faces'.format(((FP)/(FP+TN)*100)))
print('-'*50)
hlp.show_images(image_files = dogs_detected_in_faces, n_col = 4, image_size = 50, h = 10, w = 10)
plt.show();
display_performance(TP, FN, TN, FP)
We suggest VGG-16 as a potential network to detect dog images in your algorithm, but you are free to explore other pre-trained networks (such as Inception-v3, ResNet-50, etc). Please use the code cell below to test other pre-trained PyTorch models. If you decide to pursue this optional task, report performance on human_files_short and dog_files_short.
### (Optional)
### TODO: Report the performance of another pre-trained network.
### Feel free to use as many code cells as needed.
Step 3: Create a CNN to Classify Dog Breeds (from Scratch)¶
Now that we have functions for detecting humans and dogs in images, we need a way to predict breed from images. In this step, you will create a CNN that classifies dog breeds. You must create your CNN from scratch (so, you can't use transfer learning yet!), and you must attain a test accuracy of at least 10%. In Step 4 of this notebook, you will have the opportunity to use transfer learning to create a CNN that attains greatly improved accuracy.
We mention that the task of assigning breed to dogs from images is considered exceptionally challenging. To see why, consider that even a human would have trouble distinguishing between a Brittany and a Welsh Springer Spaniel.
| Brittany | Welsh Springer Spaniel |
|---|---|
 |
 |
It is not difficult to find other dog breed pairs with minimal inter-class variation (for instance, Curly-Coated Retrievers and American Water Spaniels).
| Curly-Coated Retriever | American Water Spaniel |
|---|---|
 |
 |
Likewise, recall that labradors come in yellow, chocolate, and black. Your vision-based algorithm will have to conquer this high intra-class variation to determine how to classify all of these different shades as the same breed.
| Yellow Labrador | Chocolate Labrador | Black Labrador |
|---|---|---|
 |
 |
 |
We also mention that random chance presents an exceptionally low bar: setting aside the fact that the classes are slightly imabalanced, a random guess will provide a correct answer roughly 1 in 133 times, which corresponds to an accuracy of less than 1%.
Remember that the practice is far ahead of the theory in deep learning. Experiment with many different architectures, and trust your intuition. And, of course, have fun!
(IMPLEMENTATION) Specify Data Loaders for the Dog Dataset¶
Use the code cell below to write three separate data loaders for the training, validation, and test datasets of dog images (located at dog_images/train, dog_images/valid, and dog_images/test, respectively). You may find this documentation on custom datasets to be a useful resource. If you are interested in augmenting your training and/or validation data, check out the wide variety of transforms!
Default values¶
batch_size = 20
image_size = 315
base_lr = 0.0065 # used by get_triangular_lr for generating learning rate
max_lr = 0.004 # used by get_triangular_lr for generating learning rate
optim_type = 'Adam'
model_scratch_save_path = 'model_scratch.pt'
Load and transform Data¶
Following are functions for finding and handling class imbalance, if it exist.
get_class_info returns indexes of classes with minimum and maximum count.
create_weighted_random_sampler creates a WeightedRandomSampler which can be passed as sampler=WeightedRandomSampler to DataLoader when there is a class imbalance.
get_class_weights returns class_weights, which is then passed to CrossEntropy. This option can also be used during class imbalance
def get_class_info(data):
classes = len(data.classes)
samples={}
samples = dict(collections.Counter(sample_tup[1] for sample_tup in data.imgs))
sorted_x = sorted(samples.items(), key=operator.itemgetter(1))
min_class_idx = sorted_x[0][0]
n_min_class = sorted_x[0][1]
max_class_idx = sorted_x[classes-1][0]
n_max_class = sorted_x[classes-1][1]
return min_class_idx, n_min_class, max_class_idx, n_max_class
def get_sample_counts(data):
samples = dict(collections.Counter(sample_tup[1] for sample_tup in data.imgs))
return list(val for val in samples.values())
# below function creates WeightedRandomSampler
if torch.__version__ == '1.0.0':
from torch.utils.data import Dataset, WeightedRandomSampler, DataLoader
def create_weighted_random_sampler(data, replacement = True):
class_sample_counts = get_sample_counts(data)
weights = 1. / torch.tensor(class_sample_counts, dtype=torch.float)
sample_targets = list(sample_tup[1] for sample_tup in data_scratch['train'].imgs)
samples_weights = weights[sample_targets]
return WeightedRandomSampler( weights=samples_weights, num_samples=len(samples_weights), replacement=replacement)
# below function creates class_weights to be passed to CrossEntropty
def get_class_weights(data):
class_sample_counts = get_sample_counts(data)
class_weights = 1. / torch.tensor(class_sample_counts, dtype=torch.float)
return class_weights
Follwoing function return's 3 different types of transformations. None is used by valid and test loaders, while 'Default is used by train loader. 'Visualize' is same as Default type, except it does not add ToTensor and Normalize transformation.
def get_image_transforms(image_size, trans_type='Default'):
# Valid trans_type
assert trans_type==None or trans_type =='Default' or trans_type == 'Visualize'
# Crop size
out_size = math.floor(image_size*0.9)
img_transforms=[]
if trans_type=='Default' or trans_type=='Visualize':
img_transforms.append(transforms.RandomHorizontalFlip())
img_transforms.append(transforms.RandomRotation(30))
img_transforms.append(transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4, hue=0.1))
img_transforms.append(transforms.RandomGrayscale(p=0.4))
img_transforms.append(transforms.RandomResizedCrop(out_size, scale=(0.08,1), ratio=(1,1)))
else:
img_transforms.append(transforms.CenterCrop(out_size))
if not trans_type=='Visualize':
img_transforms.append(transforms.ToTensor())
img_transforms.append(transforms.Normalize((0.485, 0.456, 0.406),(0.229, 0.224, 0.225)))
return img_transforms
load_images funtion creates 3 DataLoader i.e. train/valid/test for image transformations during learning and testing phase.
import os
import torch
from torchvision import datasets,transforms
from PIL import ImageFile
import collections
import operator
ImageFile.LOAD_TRUNCATED_IMAGES = True
## Specify appropriate transforms, and batch_sizes
def load_images(num_workers=0, batch_size=batch_size, image_size = image_size, trans_type='Default'):
data_dir = 'data/dogImages_crop_350/'
train_dir = os.path.join(data_dir, 'train/')
valid_dir = os.path.join(data_dir, 'valid/')
test_dir = os.path.join(data_dir, 'test/')
# load and transform data using ImageFolder
train_transforms = get_image_transforms(image_size, trans_type)
data_train_transform = transforms.Compose(np.array(train_transforms))
none_transforms = get_image_transforms(image_size, trans_type=None)
data_none_transform = transforms.Compose(np.array(none_transforms))
train_data = datasets.ImageFolder(train_dir, transform=data_train_transform)
valid_data = datasets.ImageFolder(valid_dir, transform=data_none_transform)
test_data = datasets.ImageFolder(test_dir, transform=data_none_transform)
min_class_idx, n_min_class, max_class_idx, n_max_class = get_class_info(train_data)
trainloader = torch.utils.data.DataLoader(train_data, batch_size=batch_size, shuffle=True,\
num_workers=num_workers)
validloader = torch.utils.data.DataLoader(valid_data, batch_size=batch_size, num_workers=num_workers)
testloader = torch.utils.data.DataLoader(test_data, batch_size=batch_size, num_workers=num_workers)
data_scratch = dict(train=train_data, valid=valid_data, test=test_data)
loaders_scratch = dict(train=trainloader, valid=validloader, test=testloader)
return data_scratch, loaders_scratch, min_class_idx, n_min_class, max_class_idx, n_max_class
data_scratch, loaders_scratch, min_class_idx, n_min_class, max_class_idx, n_max_class = load_images()
use_cuda = torch.cuda.is_available()
n_classes = len(data_scratch['train'].classes)
# print out some data stats
print('No# training images: ', len(data_scratch['train']))
print('No# validation images: ', len(data_scratch['valid']))
print('No# test images: ', len(data_scratch['test']))
print('No# test classes: ', n_classes)
print('Cuda enabled:', use_cuda)
if n_min_class == n_max_class:
print('\nSamples in each class: {}'.format(n_min_class))
else:
print('\nClass with minimum samples:- {}({} samples)'.format(data_scratch['train'].classes[min_class_idx], n_min_class))
print('Class with maximum samples:- {}({} samples)'.format(data_scratch['train'].classes[max_class_idx], \
n_max_class))
Question 3: Describe your chosen procedure for preprocessing the data.
- How does your code resize the images (by cropping, stretching, etc)? What size did you pick for the input tensor, and why?
- Did you decide to augment the dataset? If so, how (through translations, flips, rotations, etc)? If not, why not?
Answer:¶
Applied Data augmentation using torchvision, as details are mentioned below. The model starts overfitting after 3rd epoch, if augmentation is not applied.
Data augmentation using
cv2:Below augmenation applied on training data before being fed to the model:
Resizedthe images to 350 x 350. Preserved the aspect ration while resizingCentreCropof size (315) x (315) i.e. 350x0.9=315
Using
cv2, resized and croped the image files to 315x315 during testings. Also it reduced the zip size to358 MBfrom1.05 GB
- Data augmentation using
torchvision:- Following augmenation applied on training data:
- RandomHorizontalFlip()
- RandomRotation(30)
- ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4, hue=0.1)
- RandomGrayscale(p=0.4)
- RandomResizedCrop(
283, scale=(0.08,1), ratio=(1,1)) - Normalize((0.485, 0.456, 0.406),(0.229, 0.224, 0.225))
- Following augmenation applied on training data:
- Following augmenation applied on validation and test data:
- CenterCrop(
283) - Normalize((0.485, 0.456, 0.406),(0.229, 0.224, 0.225))
- CenterCrop(
- Input size to Model
283x283i.e.315 x 0.9=283[0.9is the ratio for cropping]. The image gets cropped twice, i.e once during resize and CenterCrop transformation (size: from350x350to315x315) using cv2 and next in RandomResizedCrop (size: from315x315to283 x283) during training using torchvision.
- Reason to use a slightly larger input size (i.e 283) compared to other well known pre-trained networks was,
- To increase number of parameters given to the
AdaptiveAvgPool2d(output_size=8). If input size is 268x268, then input toAdaptiveAvgPool2dis 9x9, where as for input size 283x283 to network, the input toAdaptiveAvgPool2dis 10x10 - Also not using a larger size than 350x350 was to avoid memory issues.
- To increase number of parameters given to the
There is a class imbalance in dataset, i.e. Norwegian buhund class has 26 samples while Alaskan malamute has 77 samples. Using pytorch, there are 2 approaches to handle class imbalance:
We can use
WeightedRandomSamplerand assign it tosamplerin dataloader. ***But works on version 1.0.0 and above.
We can also assign class
weightstoCrossEntropyLossi.e. assign large weight to under-sampled classes and less weight to over-sampled classes
[Used 2nd option for class imbalance during testing. But didn't see any performance gain. Also saw increase in loss and processing time.]
(IMPLEMENTATION) Model Architecture¶
Create a CNN to classify dog breed. Use the template in the code cell below.
""" F I N A L M O D E L from Test-7 [Also defined in models.py]"""
class Net(nn.Module):
# same as Net_3 except for in_features
def __init__(self):
super().__init__()
#self.max_pool0 = nn.MaxPool2d(kernel_size=2, stride=2)
self.conv2d1_1 = nn.Conv2d(3, 32, kernel_size = 3, stride = 1, bias=False)
self.pool1_1 = nn.MaxPool2d(kernel_size=3, stride=2)
self.conv2d1_2 = nn.Conv2d(32, 32, kernel_size = 3, stride = 1, padding = 1,bias=False)
self.pool1_2 = nn.MaxPool2d(kernel_size=3, stride=2)
self.BatchNorm2d1 = nn.BatchNorm2d(32)
self.LeakyReLU1 = nn.LeakyReLU(negative_slope=0.1)
self.conv2d2_1 = nn.Conv2d(32, 64, kernel_size = 3, stride = 1, bias=False)
self.pool2_1 = nn.MaxPool2d(kernel_size=3, stride=2)
self.conv2d2_2 = nn.Conv2d(64, 64, kernel_size = 3, stride = 1, padding = 1,bias=False)
self.pool2_2 = nn.MaxPool2d(kernel_size=3, stride=2)
self.BatchNorm2d2 = nn.BatchNorm2d(64)
self.LeakyReLU2 = nn.LeakyReLU(negative_slope=0.1)
self.conv2d3_1 = nn.Conv2d(64, 128, kernel_size = 3, stride = 1, bias=False)
self.pool3_1 = nn.MaxPool2d(kernel_size=3, stride=1)
self.conv2d3_2 = nn.Conv2d(128, 128, kernel_size = 3, stride = 1, padding = 1,bias=False)
self.pool3_2 = nn.MaxPool2d(kernel_size=3, stride=1)
self.BatchNorm2d3 = nn.BatchNorm2d(128)
self.LeakyReLU3 = nn.LeakyReLU(negative_slope=0.1)
self.avg_pool = nn.AdaptiveAvgPool2d(output_size=8)
self.fc1 = nn.Linear(in_features=8192, out_features=1024, bias=False)
self.fc1_bn1d = nn.BatchNorm1d(1024)
self.fc1_relu = nn.ReLU()
self.fc2 = nn.Linear(in_features=1024, out_features=n_classes, bias=False)
self.fc2_bn1d = nn.BatchNorm1d(n_classes)
def forward(self, x):
x = self.LeakyReLU1(self.BatchNorm2d1(self.pool1_2(self.conv2d1_2(self.pool1_1(self.conv2d1_1(x))))))
x = self.LeakyReLU2(self.BatchNorm2d2(self.pool2_2(self.conv2d2_2(self.pool2_1(self.conv2d2_1(x))))))
x = self.LeakyReLU3(self.BatchNorm2d3(self.pool3_2(self.conv2d3_2(self.pool3_1(self.conv2d3_1(x))))))
x = self.avg_pool(x)
x = x.view(-1, 8192)
x = self.fc1_relu(self.fc1_bn1d( self.fc1( x ) ) )
x = self.fc2_bn1d( self.fc2( x ) )
return x
#-#-# You so NOT have to modify the code below this line. #-#-#
""" instantiate the CNN """
model_scratch = Net()
""" move tensors to GPU if CUDA is available """
if use_cuda:
model_scratch.cuda()
Weight initialization¶
def weights_init(m):
'''Takes in a module and initializes all linear layers with weight
values taken from a normal distribution.'''
if isinstance(m, nn.Conv2d) or isinstance(m, nn.Linear):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
model_scratch.apply(weights_init)
Visualizing feature maps¶
- Below function shows the image transformation steps applied by torchvision during training.
- Also below we can see, randomly selected feature maps generated by convolutional layer
image_transforms = get_image_transforms(image_size, trans_type = 'Visualize')
hlp.visualize_model_cnn_output(image_size, model_scratch, image_transforms=image_transforms) #194-228
Question 4: Outline the steps you took to get to your final CNN architecture and your reasoning at each step.
Answer:
- I took initial model from Lesson-1, 37-Notebook: CNNs for CIFAR Image Classification. I performed around 21 different tests, with 10 epochs and test-accuracy between 10% - 20%. The details of test models used can be seen in models.py. The final model was derived from the test model
Net_21. All the models stopped improving after 20 epochs.
- The final model takes some inspiration from VGG network.
- Convolutional and Linear layers are initialzed using kaiming normalization. BatchNorm2d weights are initialized to 1 and bias is initialize to 0.
- The model has 3 sets of convolutional layers.
- In each layer, there are 2 convolution operations, 2 max pool operations , 1 Batch normaliztion and in last Leaky ReLU activation.
- I pass the data to fully-connected layer using
AdaptiveAvgPool2d. - After adding
AdaptiveAvgPool2dand I removedDropoutfrom the fully-connected layer. I saw improvement in performance from 35% to upto 50%.
Learning rate: I have used The Cyclical Learning Rate technique, which improved the performance while training the model instead of fixed learning rate. I start with a high learning rate, i.e
base_lr = 0.0065and decreasing to low learning rate uptomax_lr = 0.004. As the batch_size is20, I reset the learning rate after 20 epochs while training. e.g. when batch_size is5:Normal Cyclic LR Reset Cyclic LR -------------------- -------------------- epoch: 0 Lr: 0.0065 epoch: 0 Lr: 0.0065 epoch: 1 Lr: 0.006 epoch: 1 Lr: 0.006 epoch: 2 Lr: 0.0055 epoch: 2 Lr: 0.0055 epoch: 3 Lr: 0.005 epoch: 3 Lr: 0.005 epoch: 4 Lr: 0.0045 epoch: 4 Lr: 0.0045 epoch: 5 Lr: 0.004 [epoch: 5 Lr: 0.0065] [epoch: 6 Lr: 0.0045] epoch: 6 Lr: 0.006 epoch: 7 Lr: 0.005 epoch: 7 Lr: 0.0055 epoch: 8 Lr: 0.0055 epoch: 8 Lr: 0.005 epoch: 9 Lr: 0.006 epoch: 9 Lr: 0.0045We can see that incase of
normal Cyclic LR, after epoch-5, the LR starts increasing, While incase ofReset Cyclic LR, after epoch-4, the LR restarts from 0.0065.
- Added
np.random.seed(42),torch.backends.cudnn.deterministic = Trueandtorch.cuda.manual_seed_all(42)for consistent training results. But still saw minor differences in the results
(IMPLEMENTATION) Specify Loss Function and Optimizer¶
Use the next code cell to specify a loss function and optimizer. Save the chosen loss function as criterion_scratch, and the optimizer as optimizer_scratch below.
import torch.optim as optim
#Ref: http://teleported.in/posts/cyclic-learning-rate/
def get_triangular_lr(iteration, stepsize, base_lr=base_lr, max_lr=max_lr):
"""Given the inputs, calculates the lr that should be applicable for this iteration"""
cycle = np.floor(1 + iteration/(2 * stepsize))
x = np.abs(iteration/stepsize - 2 * cycle + 1)
lr = base_lr + (max_lr - base_lr) * np.maximum(0, (1-x))
return lr
### TODO: select loss function
#cross_entropy_loss_weights = get_class_weights(data_scratch['train'])
def create_criterion():
#return nn.CrossEntropyLoss(cross_entropy_loss_weights)
return nn.CrossEntropyLoss()
### TODO: select optimizer
def create_optimizer(parameters, op_type=optim_type,batch_size=batch_size, epoch=1,base_lr=base_lr, max_lr=max_lr):
assert op_type == 'SGD' or op_type == 'Adam'
op_lr = get_triangular_lr(epoch, batch_size, base_lr, max_lr)
if op_type == 'SGD':
return optim.SGD(parameters,
lr=op_lr, momentum=0.9), op_lr
elif op_type == 'Adam':
return optim.Adam(parameters, lr=op_lr), op_lr
criterion_scratch = create_criterion()
optimizer_scratch, lr_scratch = create_optimizer(model_scratch.parameters(), op_type = optim_type)
(IMPLEMENTATION) Train and Validate the Model¶
Train and validate your model in the code cell below. Save the final model parameters at filepath 'model_scratch.pt'.
ImageFile.LOAD_TRUNCATED_IMAGES = True
def train(n_epochs, loaders, model, optimizer, lr, criterion, use_cuda, save_path, loss_csv, \
recs = len(data_scratch['train']), start_epoch=1):
def update_lr(lr):
for param_group in optimizer.param_groups:
param_group['lr'] = lr
# initialize tracker for minimum validation loss
valid_loss_min = np.Inf
if len(loss_csv)>0:
valid_loss_min = min(loss_csv['Validation Loss'])
model_test = copy.deepcopy(model)
# loss trackers
prev_valid_loss = np.Inf
prev_train_loss = np.Inf
print('='*127)
print('| Epoch | Training Loss | Validation Loss | Elapsed Time | optim | Model Status | Recs |')
print('='*127)
df_loss = loss_csv
loss_file = 'loss_'+save_path.replace('.pt','')+'.csv'
for epoch in range(start_epoch, n_epochs+1):
start = time.time()
# initialize variables to monitor training and validation loss
train_loss = 0.0
valid_loss = 0.0
# set RNG seeds for python
np.random.seed(42+ epoch)
# set RNG seeds for pytorch
if use_cuda:
torch.backends.cudnn.deterministic = True
torch.cuda.manual_seed_all(42+ epoch)
else:
torch.manual_seed(42+ epoch)
###################
# train the model #
###################
model.train()
"""
To be delete
============
data_images, loaders, _, _, _, _ = load_images(trans_type='Random', dir_seq=(epoch-1)%35)
recs = len(data_images['train'])
"""
for batch_idx, (data, target) in enumerate(loaders['train']):
# move to GPU
if use_cuda:
data, target = data.cuda(), target.cuda() # target.shape: [20(batch-size)]
## find the loss and update the model parameters accordingly
## record the average training loss, using something like
optimizer.zero_grad()
output = model(data) # output.shape: [(batch-size), (classes)]
loss = criterion(output, target)
loss.backward()
optimizer.step()
train_loss += loss.item()*data.size(0)
######################
# validate the model #
######################
model.eval()
for batch_idx, (data, target) in enumerate(loaders['valid']):
# move to GPU
if use_cuda:
data, target = data.cuda(), target.cuda()
## update the average validation loss
output = model(data)
loss = criterion(output, target)
valid_loss += loss.item()*data.size(0)
train_loss = train_loss/len(loaders['train'].dataset)
valid_loss = valid_loss/len(loaders['valid'].dataset)
end=time.time()
elapsed = end - start
## TODO: save the model if validation loss has decreased
model_status=''
loss_min_reduced = valid_loss < valid_loss_min
if loss_min_reduced:
torch.save(model.state_dict(), save_path)
valid_loss_min = valid_loss
model_status='Model saved'
train_loss_increased = train_loss > prev_train_loss
valid_loss_increased = valid_loss > prev_valid_loss
train_loss_reduced = train_loss < prev_train_loss
valid_loss_reduced = valid_loss < prev_valid_loss
curr_overfitting = (train_loss < valid_loss)
curr_underfit = train_loss_increased and valid_loss_increased
if curr_overfitting:
if model_status == '':
model_status = 'Model overfitting'
else:
model_status = model_status + '(overfitting)'
if curr_underfit and not curr_overfitting:
if model_status == '':
model_status = 'Model underfitting'
else:
model_status = model_status + '(underfitting)'
prev_valid_loss = valid_loss
prev_train_loss = train_loss
# print training/validation statistics
df_loss.loc[epoch]=[optim_type, train_loss, valid_loss, elapsed, lr, model_status, recs]
print('|{}|{}|{}|{}|{}|{}|{}|'.format(str(epoch).rjust(7),str(round(train_loss,10)).rjust(19),\
str(round(valid_loss,10)).rjust(19),str(round(elapsed,5)).rjust(18),\
(str(round(lr,8))+'('+optim_type+')').rjust(16) ,model_status.rjust(30),\
str(recs).rjust(10)))
lr = get_triangular_lr((epoch+1)%(batch_size+1), batch_size, base_lr, max_lr)
update_lr(lr)
df_loss.to_csv(loss_file, sep='\t')
print('='*127)
# return trained model
return model, df_loss
loss_csv = pd.DataFrame(columns=['optim','Training Loss','Validation Loss','Elapsed Time','lr','Model Saved','Recs'])
model, df_loss = train(100, loaders_scratch, model_scratch, optimizer_scratch, lr_scratch, \
criterion_scratch, use_cuda, model_scratch_save_path, loss_csv=loss_csv)
(IMPLEMENTATION) Test the Model¶
Try out your model on the test dataset of dog images. Use the code cell below to calculate and print the test loss and accuracy. Ensure that your test accuracy is greater than 10%.
# test the model
def test(loaders, model, criterion, use_cuda, verbose=True):
# monitor test loss and accuracy
test_loss = 0.
correct = 0.
total = 0.
if use_cuda:
model.cuda()
model.eval()
for batch_idx, (data, target) in enumerate(loaders['test']):
# move to GPU
if use_cuda:
torch.backends.cudnn.deterministic = True
torch.cuda.manual_seed_all(42)
data, target = data.cuda(), target.cuda()
# forward pass: compute predicted outputs by passing inputs to the model
output = model(data)
# calculate the loss
loss = criterion(output, target)
# update average test loss
test_loss = test_loss + ((1 / (batch_idx + 1)) * (loss.data - test_loss))
# convert output probabilities to predicted class
pred = output.data.max(1, keepdim=True)[1]
# compare predictions to true label
correct += np.sum(np.squeeze(pred.eq(target.data.view_as(pred))).cpu().numpy())
total += data.size(0)
if verbose:
print('='*127)
print('| Loss | {}| Test Accuracy | {} |'.format(str(round(test_loss.item(),11)).rjust(16),\
(str(100 * correct // total)+ '% ('+ str(correct)+'/'+ str(total)+')').ljust(74)))
print('='*127)
return test_loss, correct, total
# load the model that got the best validation accuracy
model_scratch.load_state_dict(torch.load(model_scratch_save_path, map_location='cpu'),strict=False)
# call test function
test_loss, correct, total = \
test(loaders_scratch, model_scratch, criterion_scratch, use_cuda)
Model Loss¶
After 61 epochs the model starts over-fitting
loss_file = 'loss_'+model_scratch_save_path.replace('.pt','')+'.csv'
dfLoss=pd.read_csv(loss_file, '\t')
#log24.columns=['idx','train_loss','valid_loss','elapsed_time','lr_no','model_status']
fig, ax = plt.subplots(figsize=(18,7))
dfLoss['Training Loss'].plot(legend=True, ax=ax);
dfLoss['Validation Loss'].plot(legend=True, ax=ax);
ax.set_xlabel('Epoch')
ax.set_ylabel('Loss')
plt.show()
Step 4: Create a CNN to Classify Dog Breeds (using Transfer Learning)¶
You will now use transfer learning to create a CNN that can identify dog breed from images. Your CNN must attain at least 60% accuracy on the test set.
(IMPLEMENTATION) Specify Data Loaders for the Dog Dataset¶
Use the code cell below to write three separate data loaders for the training, validation, and test datasets of dog images (located at dogImages/train, dogImages/valid, and dogImages/test, respectively).
If you like, you are welcome to use the same data loaders from the previous step, when you created a CNN from scratch.
batch_size = 20
base_lr = 0.001 # used by get_triangular_lr for generating learning rate
max_lr = 0.001 # used by get_triangular_lr for generating learning rate
optim_type = 'SGD'
model_transfer_save_path = 'model_transfer.pt'
## TODO: Specify data loaders
data_transfer, loaders_transfer, _, _,_,_ = load_images(batch_size = batch_size, trans_type='Default')
(IMPLEMENTATION) Model Architecture¶
Use transfer learning to create a CNN to classify dog breed. Use the code cell below, and save your initialized model as the variable model_transfer.
import torchvision.models as models
import torch.nn as nn
import copy
## TODO: Specify model architecture
model_transfer = models.resnet50(pretrained='imagenet')
for param in model_transfer.parameters():
param.requires_grad = False
model_transfer.fc = nn.Linear(18432, n_classes, bias=True)
fc_parameters = model_transfer.fc.parameters()
for param in fc_parameters:
param.requires_grad = True
print(model_transfer)
if use_cuda:
model_transfer = model_transfer.cuda()
model_test = copy.deepcopy(model_transfer)
Question 5: Outline the steps you took to get to your final CNN architecture and your reasoning at each step. Describe why you think the architecture is suitable for the current problem.
Answer:
VGG-16 and Resnet are the modern network architecture.
VGG-16when introduced in 2014 was considered one of the very deep model. It is a very simple architecture, yet very powerful. But is very slow and the weights themselves are very huge.Resnet was introduced in 2015, is a Deep residual network, build using micro-architecture modules( also called building blocks). These
Residual blocksintroduced by the authors of Resnet, helped in addressing the modelDegradtionproblem. It is much faster compared toVGG-16I chose
Resnet50to build the final CNN architecture for transfer learning. I had to replace the fully connected layer withLinear(in_features=18432, out_features=133, bias=True).Also we have to pass fc.parameters to the optimizer as its going to learn the classifier weights.
(IMPLEMENTATION) Specify Loss Function and Optimizer¶
Use the next code cell to specify a loss function and optimizer. Save the chosen loss function as criterion_transfer, and the optimizer as optimizer_transfer below.
criterion_transfer = create_criterion()
optimizer_transfer, lr_transfer = \
create_optimizer(model_transfer.fc.parameters(), op_type=optim_type, batch_size=batch_size, epoch=1, base_lr=base_lr, max_lr=max_lr)
(IMPLEMENTATION) Train and Validate the Model¶
Train and validate your model in the code cell below. Save the final model parameters at filepath 'model_transfer.pt'.
loss_csv = pd.DataFrame(columns=['optim','Training Loss','Validation Loss','Elapsed Time','lr','Model Saved','Recs'])
model, df_loss = train(10, loaders_transfer, model_transfer, optimizer_transfer, lr_transfer, \
criterion_transfer, use_cuda, model_transfer_save_path, loss_csv=loss_csv)
(IMPLEMENTATION) Test the Model¶
Try out your model on the test dataset of dog images. Use the code cell below to calculate and print the test loss and accuracy. Ensure that your test accuracy is greater than 60%.
# load the model that got the best validation accuracy
model_transfer.load_state_dict(torch.load(model_transfer_save_path, map_location='cpu'),strict=False)
# call test function
test_loss, correct, total = \
test(loaders_transfer, model_transfer, criterion_transfer, use_cuda)
Model Loss¶
loss_file = 'loss_'+model_transfer_save_path.replace('.pt','')+'.csv'
dfLoss=pd.read_csv(loss_file, '\t')
#log24.columns=['idx','train_loss','valid_loss','elapsed_time','lr_no','model_status']
fig, ax = plt.subplots(figsize=(18,7))
dfLoss['Training Loss'].plot(legend=True, ax=ax);
dfLoss['Validation Loss'].plot(legend=True, ax=ax);
ax.set_xlabel('Epoch')
ax.set_ylabel('Loss')
plt.show()
model_transfer.load_state_dict(torch.load(model_transfer_save_path, map_location='cpu'),strict=False)
(IMPLEMENTATION) Predict Dog Breed with the Model¶
Write a function that takes an image path as input and returns the dog breed (Affenpinscher, Afghan hound, etc) that is predicted by your model.
### TODO: Write a function that takes a path to an image as input
### and returns the dog breed that is predicted by the model.
from PIL import Image
# list of class names by index, i.e. a name can be accessed like class_names[0]
class_names = [item[4:].replace("_", " ") for item in data_transfer['train'].classes]
def predict_breed_transfer(img_path, model, use_cuda):
image = Image.open(img_path).convert('RGB')
out_size = math.floor(image_size*0.9)
test_transforms = get_image_transforms(image_size, trans_type=None)
image_transform = transforms.Compose(np.array(test_transforms))
data = image_transform(image)
data = data.unsqueeze_(0)
if use_cuda:
data = data.cuda()
model.eval()
output = model(data)
probas = F.softmax(output.data.cpu(), dim=1)*100
return probas
Step 5: Write your Algorithm¶
Write an algorithm that accepts a file path to an image and first determines whether the image contains a human, dog, or neither. Then,
- if a dog is detected in the image, return the predicted breed.
- if a human is detected in the image, return the resembling dog breed.
- if neither is detected in the image, provide output that indicates an error.
You are welcome to write your own functions for detecting humans and dogs in images, but feel free to use the face_detector and human_detector functions developed above. You are required to use your CNN from Step 4 to predict dog breed.
Some sample output for our algorithm is provided below, but feel free to design your own user experience!
(IMPLEMENTATION) Write your Algorithm¶
def get_embeded_image(img_path):
data_uri = base64.b64encode(open(img_path, 'rb').read()).decode('utf-8').replace('\n', '')
return '<img src="data:image/png;base64,{0}" height=350 width=250>'.format(data_uri)
def ShowResult(name_from_file, imagenet_label, human_name, dog_detected, breed, breed_strenght, probas, img_path):
human_face_detected = '{}'.format(human_name is not None)
human_name = '{}'.format(human_name if human_name is not None else '-')
dog_detected = '{}'.format(dog_detected)
dog_breed = '{}{}'.format(breed if breed is not None else '-', \
'( '+str(breed_strenght)+'% )' if breed is not None else '')
row_start = '<tr><td><font size="3">'
row_middle = '</font></td><td><b><font size="3"><p style="text-align:left;">'
row_end = '</p></font></b></td></tr>'
display(HTML('<table><tr><td width="60%"><table>\
<tr><td colspan="2" ><h1 style="text-align:center;"><b>'+name_from_file.ljust(50, ' ')+'</b></h1></td></tr>\
'+row_start+'Imagenet Label:' + row_middle + imagenet_label + row_end +'\
'+row_start+'Face Detected:' + row_middle + human_face_detected + row_end +'\
'+row_start+'Name: ' + row_middle + human_name + row_end +'\
'+row_start+'Dog Detected: ' + row_middle + dog_detected + row_end +'\
'+row_start+'Breed: ' + row_middle + dog_breed + row_end +'\
</td></tr></table></td><td width="40%">' + get_embeded_image(img_path) + '</td></tr></table>'))
### TODO: Write your algorithm.
### Feel free to use as many code cells as needed.
def process_face_detector(img_path):
face_detected = face_detector(img_path)
name = None
if face_detected:
name = hlp.get_name_from_filename(img_path)
return name
def process_dog_detector(img_path):
dog_detected = dog_detector(img_path)
breed = None
breed_strenght = None
probas = None
if dog_detected:
model_transfer.load_state_dict(torch.load(model_transfer_save_path, map_location='cpu'),strict=False)
probas = predict_breed_transfer(img_path, model_transfer, use_cuda)
breed = class_names[probas.argmax()]
breed_strenght = round(probas.max().item(),0)
return dog_detected, breed, breed_strenght, probas
def run_app(img_path):
## handle cases for a human face, dog, and neither
name_from_file = hlp.get_name_from_filename(img_path)
imagenet_label = predict_imagenet_class_label(img_path)
human_name = process_face_detector(img_path)
dog_detected, breed, breed_strenght, probas = process_dog_detector(img_path)
ShowResult(name_from_file, imagenet_label, human_name, dog_detected, breed, breed_strenght, probas, img_path)
Step 6: Test Your Algorithm¶
In this section, you will take your new algorithm for a spin! What kind of dog does the algorithm think that you look like? If you have a dog, does it predict your dog's breed accurately? If you have a cat, does it mistakenly think that your cat is a dog?
(IMPLEMENTATION) Test Your Algorithm on Sample Images!¶
Test your algorithm at least six images on your computer. Feel free to use any images you like. Use at least two human and two dog images.
Question 6: Is the output better than you expected :) ? Or worse :( ? Provide at least three possible points of improvement for your algorithm.
Answer:
- Fine tune the resnet transfer learning model, by trying different optimizers and learning rates. Currently used only
SGDwithLr=0.001. Also should test the performane usingF1/F2-scores
- Fine tune the resnet transfer learning model, by trying different optimizers and learning rates. Currently used only
- Even the data used in transfer learning was not balanced. Add more sample augmented images, using
openCV,imgaug,skimage
- Even the data used in transfer learning was not balanced. Add more sample augmented images, using
- Also try other pre-trained architectures like
Inception v3,ResNet-152,VGG-19
- Also try other pre-trained architectures like
## TODO: Execute your algorithm from Step 6 on
## at least 6 images on your computer.
## Feel free to use as many code cells as needed.
##
## suggested code, below
human_samples = random.sample(range(0, len(human_files)-1), 6)
dog_samples = random.sample(range(0, len(dog_files)-1), 6)
for file in np.hstack((human_files[human_samples], dog_files[dog_samples])):
run_app(file)
|  | ||||||||||||
|  | ||||||||||||
|  | ||||||||||||
|  | ||||||||||||
|  | ||||||||||||
|  | ||||||||||||
|  | ||||||||||||
|  | ||||||||||||
|  | ||||||||||||
|  | ||||||||||||
|  | ||||||||||||
|  | ||||||||||||